Road Traffic Volume Prediction And Environmental Impact Research Based On Deep Learning

With the increase of vehicle ownership in China,the air pollution caused by vehicle emission becomes more and more serious.In order to grasp the traffic flow status and alleviate traffic pollution,it is necessary to predict road traffic volume and study the environmental impact.In the existing research on traffic volume prediction,fixed weight is often used to represent the relationship between prediction nodes and adjacent nodes,and the relationship between nodes in the spatial and temporal dimensions is seldom considered at present.In the research on the impact of traffic on air pollution,most studies focus on analyzing the correlation between pollution concentration and traffic volume.Few studies have explored the lag effect between the two.In the prediction of pollution concentration,the problem of incomplete extraction of spatiotemporal characteristics of pollution concentration at multiple monitoring points still exists.Based on this,the traffic pollution monitoring points are taken as the research area.The time delay quantitative study of road traffic emissions on traffic pollution concentration is carried out.Then,a traffic volume prediction model based on road network space time map with improved weight design is proposed,which integrates the spatio temporal characteristics of traffic flow.Later,a traffic pollution concentration prediction model based on the spatio temporal characteristics of multiple stations is proposed to fully explore the spatio temporal characteristics of pollution concentration.Finally,on the basis of the above research,an integrated prediction model of traffic pollution concentration is designed to accurately predict traffic volume and traffic pollution concentration.The main research work and conclusions are as follows:（1）The time delay effect of road traffic emissions on traffic pollution concentration is investigated.The lag effect is tested and the delay value range of different pollutants is determined based on the significant period analysis method of wavelet transform and the delay value analysis method of leading lag correlation coefficient.The reliability of the time delay results is verified by the variation of the average wavelet energy in the frequency band near the significant period.The results show that the significant period of traffic volume and traffic pollution concentration is 24h.The delay values of NO₂ and CO are 5h,6h or 7h,PM2.5 is 5h,and O₃ is 3h or 4h.（2）A traffic volume prediction model WTG CRNN based on spatio temporal graph of road network is proposed.By constructing the spatiotemporal map of the traffic network and defining the relative proximity to improve the representation of weight between nodes,the model makes the design of weight between nodes more flexible.In this model,the convolution operation is used to extract the temporal and spatial features,and the gated cyclic network is used to mine the temporal and spatial dependencies,which can effectively mine the temporal and spatial features of traffic flow.The results show that compared with ARIMA,LSTM and combined models like DCRNN and STG CRNN,this model has the highest R² in both training set and test set.The average improvement rate of RMSE,MAE and MAPE is 17.8%.（3）A traffic pollution concentration prediction model S CNN LSTM based on the spatio temporal characteristics of multiple stations and a traffic pollution concentration integration prediction model are proposed.The spatial and temporal features of multiple stations are designed and developed into two dimensional structures of the data of each station.The convolutional neural network is used to extract the spatial features and the long and short time memory network is used to process the time series.The spatial and temporal features of traffic pollution concentration are fully extracted.Based on WTG CRNN and S CNN LSTM,an integrated prediction model of traffic pollution concentration is established.The results show that compared with MLR,LSTM,S CNN,S LSTM and CNN LSTM,the average improvement rate of each index of S CNN LSTM is 20.8%.The advantage of the integrated prediction model in feature selection is verified by analyzing the prediction results without and with time delay.